Sparse weight mapping and computation reuse strategy for scalable photonic matrix multiplication.

Area of Science:

• Photonics
• Optical Computing
• Materials Science

Background:

• Photonic crossbar arrays using phase-change materials (PCMs) offer high integration density for parallel photonic matrix multiplication.
• Scalability of these arrays is hindered by optical transmission loss, limiting practical large-scale applications.

Purpose of the Study:

• To propose and validate a weight mapping strategy for efficient execution of larger convolutional computations within scale-limited photonic crossbar arrays.
• To address the scalability limitations imposed by optical transmission loss in photonic matrix multiplication.

Main Methods:

• Fabrication of a high-quality 4x4 photonic crossbar array with 3-bit precision modulation.
• Development and application of a novel weight mapping strategy to encode convolutional operators.
• Integration of the strategy into a photonic convolutional neural network for image processing tasks.

Main Results:

• The mapping strategy enabled efficient execution of four different 3x3 operators on the 4x4 array, yielding a 225% improvement in computational efficiency for an edge detection task.
• A photonic convolutional neural network utilizing this strategy achieved 96.7% classification accuracy on the MNIST dataset, closely matching the simulated 96.84% accuracy.

Conclusions:

• The proposed weight mapping strategy effectively overcomes hardware constraints in scale-limited photonic crossbar arrays.
• This work advances the development of large-scale photonic matrix multiplication and photonic computing by enabling efficient computation under hardware limitations.